Model Identification And Human-robot Interaction Control Of Cobots

There are two problems confronting human-robot collaboration.One is to ensure the safety when a human coexists with a robot.The other is to ensure the efficiency when a human collaborates with a robot.The present thesis deals with these problems based on model identification,disturbance observer and admittance control so that they can be solved without using exteroceptive sensors.To this end,the thesis contributes in four aspects,i.e.,static model analysis and identification,dynamic model identification,disturbance observer design and human-robot interaction control.As zero-gravity mode typically requires a static model to compensate a robot's gravity and Coulomb's friction in real time so that it can hang itself anywhere in the workspace,a static model is a prerequisite.The thesis presents a systematic identification method to obtain the static model which contains links' mass,first order mass moments and Coulomb's frictions at each joint.To be specific,the static equations can be derived by projecting the gravity of each child link onto the joint axis of the current link plus the Coulomb's friction at the joint.Separating model parameters from the equations will lead to a regressor which is often rank deficient.In order to apply the least squares,the maximum linear independent set of the regressor's columns must be determined,which is done by take advantage of the property of a revolute joint.Experiment results indicated that the proposed method can accurately identify the model parameters among which the relative errors of gravity parameters are lower than 5%.Since disturbance observer design always requires a dynamic model,the thesis proposes an identification framework to obtain a dynamic model which contains links' mass,first order mass moments,inertia matrices and friction parameters.The framework consists of three loops.In the inner loop,physical feasibility constraints are put on the model parameters to make them physically realizable.In the middle loop,outliers are removed by the iteratively reweighted least squares so that the results are robust to these outliers.In the outer loop,the nonlinear friction model is fitted to the estimated frictions to make the results accurate.In addition,an excitation trajectory optimization method is proposed so that the inertial parameters,the gravity parameters and the friction parameters can all be identified accurately.Experiment results showed that the proposed framework outperforms traditional identification methods(like WLS)in accuracy.An disturbance observer is commonly needed to provide estimated force for the interaction control.In the generalized momentum based disturbance observer,a frequently met problem is that one has to compute the derivative of the inertial matrix.To deal with this problem,the thesis proposes a recursive method to compute it based on automatic differentiation.Simulation results showed that the efficiency of the proposed method is three times of that of the method proposed by De Luca for 6 Do F serial manipulator.Another advantage of the proposed method is that it can be applied directly to an identified dynamic model.To deal with the problem that the output of a disturbance observer undergoes abrupt changes when the joint velocity cross zero,a variable bandwidth filter is proposed which have a high bandwidth when the joint speed is high and a low bandwidth when the joint speed is low.Experiment results showed that the performance of the constructed disturbance observer is limited in accuracy while it did run the same trend as the exerted forces/torques.The above mentioned static model and disturbance observer are applied in human-robot interaction control.On the one hand,to deal with the problem that the outputs of a disturbance observer can drift due to the friction with changing temperature,a dynamic threshold is set in collision detection.Experiment results showed that robots are able to detect collisions successfully with a minimum force of 20 N exerted on the end effector.On the other hand,two different solutions are provided for kinesthetic teaching(or manual guidance).One is based on the static model,the other is based on the disturbance observer with admittance control.The former consumes less computation resources and is always stable,but it offers little comfort and task consistency which make it only suitable for points teaching.On the Contrary,the latter consumes more computation resources and can be unstable,but it offers more comfort and task consistency which make it suitable for both points teaching and trajectories teaching.Simulation results showed that,in the admittance control,increasing the desired inertia,decreasing the desired damping or the filter's cutoff frequency can drive the system unstable.The experiment results showed that the disturbance observer based kinesthetic teaching outperforms the static model based kinesthetic teaching in comfortableness and task consistency.